DeepSeek, the Chinese company that has released multiple low-cost, high-performance models, has released their first model harness as a developer preview.
A harness is basically software that makes a model usable (think user interface, tools, memory, file system access, etc.) Common similar harnesses that you might have used include Claude Cowork, ChatGPT Work, or Gemini Spark.
DeepSeek’s version is still in its early stages but has a lot of promise. They are taking a different approach. The harness is modular and each part can be enabled or disabled whenever you want. As they say on their site “Everything is a plugin”. This also means that you can develop and add your own plugins or use plugins that others have created (at your own risk of course).
Their harness also includes a trajectory that allows you to look at the details of every “run” after you put in a prompt. This includes system prompts, tools used, results, sub-agent calls, context injection. Basically, everything that happens after you submit your prompt until you get output back or work is done.
Currently the harness is only version 0.1 which means that it will likely change quite a bit before it is released to the general public. This is a good thing because it is already really good.
With all this competition in the AI space, the customer will (hopefully) benefit in the end because all of these companies are keeping each other on their toes. Time will tell I guess.
If you want to take a look at DeepSeek harness here is the site - https://www.deepseek.com/harness/en/
(Probably not recommended if you aren’t a developer, or at least a tinkerer, but choose your path…)